LLM Token Counter: Practical Tips and Common Mistakes
Reviewed by the OnlineFree.app team · Updated
Key points
- The LLM Token Counter estimates tokens and input cost in the browser, with no sign-up and no stored history.
- Cost equals token count divided by one million, times the selected model preset's input price.
- The four-characters-per-token rule understates Chinese, Japanese and Korean text by roughly three to four times.
- Output tokens, conversation history and tool schemas are not counted unless you paste them into the box.
- Verify final budgets with the provider's own tokenizer, since this tool is a fast approximation.
What the LLM Token Counter measures
The LLM Token Counter estimates how many tokens a pasted prompt or document will consume and what that text costs as input, using a model preset you choose and a fast character-based heuristic rather than a full BPE tokenizer. It is built for order-of-magnitude awareness when writing prompts and sizing context, not for exact billing reconciliation.
In practice you paste text, pick a preset such as GPT-4o, Claude Sonnet or Gemini Flash, and the numbers refresh roughly 150 ms after you stop typing. There is no button to press, no sign-up and no history. The calculation runs inside the browser tab, so the text itself never leaves your machine.
The result card shows four things: the estimated token count, the character and word counts, the percentage of the selected model's context window you have consumed, and the estimated input cost in USD. Choosing the generic preset collapses the estimate to roughly one token per four characters and shows a dash instead of a price.
How the token estimate is calculated
The heuristic counts Latin-script text at roughly one token per four characters, treats CJK characters as closer to one token each, and applies separate weighting to code and punctuation. That weighting matters because real tokenizers split digits, brackets, quotes and rare Unicode sequences into many small pieces rather than merging them into long tokens.
The consequence is that two documents with the same meaning can land far apart. A 20,000-character English passage lands near 5,000 tokens, while 20,000 Chinese characters land near 20,000 tokens — roughly four times the count for similar visible length. Emoji and unusual symbols are usually worse still, since they often cost several tokens per glyph.
As of 2026 the tool deliberately does not load a real tokenizer vocabulary, which would add one to two megabytes and slow the first paint. The trade is honest: you get an answer in milliseconds that is close enough for budgeting and context-window planning, not a byte-exact figure.
Choosing a model preset before trusting the cost
Cost is computed as token count divided by one million, multiplied by the preset's input price. The built-in table lists GPT-4o at $2.50 per million input tokens, GPT-4o mini at $0.15, GPT-4.1 at $2.00, Claude Sonnet at $3.00, Claude Opus at $15.00 and Gemini Flash at $0.075. Those are static reference values and are labelled as subject to change.
The spread is large enough to change design decisions. A 10,000-token prompt costs about $0.025 on GPT-4o, $0.0015 on GPT-4o mini, $0.15 on Claude Opus and $0.00075 on Gemini Flash. The same prompt sent 500 times in a day is a few cents on the cheap presets and tens of dollars on the expensive one.
Quote the numbers with the model name attached, because context windows differ too. Claude Sonnet and Claude Opus cap at 200,000 tokens, GPT-4o and GPT-4o mini at 128,000, while GPT-4.1 and Gemini Flash reach 1,000,000. A 100,000-token document is 78% of a GPT-4o window but only 10% of a Gemini Flash window.
How many tokens does a long document need?
A 250,000-character English document estimates to about 62,500 tokens, which is roughly 49% of a 128,000-token GPT-4o context window and around $0.156 as input. Paste the same file into a 1,000,000-token preset and the percentage drops to about 6%, which is usually the point where multi-document prompting becomes practical.
Watch the input ceiling: anything past 500,000 characters is truncated with a visible notice, so a very large file will report a smaller count than the file actually contains. If you are converting a Markdown brief into a JSON payload for an API call, run it through the Markdown to JSON Converter first and then count the JSON — quotes, braces and escaped newlines all add tokens the original Markdown did not have.
Do the same when you assemble a system prompt, few-shot examples and prior conversation turns. Those blocks are easy to forget, and a 2,000-token instruction block tacked onto every request quietly multiplies across thousands of calls.
Common mistakes that skew your token count
The most frequent error is counting words and assuming tokens follow. English averages about 1.3 to 1.5 tokens per word, while code, URLs, IDs and JSON keys can run far higher because they fragment into short pieces. A 900-word prompt is often closer to 1,300 tokens than to 900.
A close second is applying the four-characters-per-token rule to non-Latin scripts. Chinese, Japanese and Korean text is far denser in tokens per character, so budgets built on the Latin ratio can be understated by three to four times. Mixed documents need both rules, which is exactly what the estimator does when it weights scripts separately.
Finally, remember what is not measured. Output tokens are excluded on purpose, since output length is a guess until the model answers, and output pricing is usually higher than input pricing. Multi-turn history, retrieved context and tool schemas are also outside the pasted text unless you paste them in. Prices in a static table can go stale, so treat the dollar figure as a planning estimate.
When to verify against a real tokenizer
Use the LLM Token Counter while you iterate — trimming a prompt, comparing two drafts, or checking whether a document fits a window — and switch to the provider's own tokenizer before you sign off on a budget or a hard context limit. Provider tokenizers are the ground truth for their models; this estimator is the fast approximation that gets you to the right order of magnitude first.
OpenAI publishes a tokenizer playground you can use for exact counts, and the Unicode Standard explains why scripts outside the Latin alphabet break the four-characters-per-token assumption. When a number matters, count it twice: once here for speed, once there for certainty.
If you only need a rough sense of scale across several presets, the free tools at OnlineFree.app cover this and other quick calculations without accounts or uploads. None of them replace a formal billing report, and none of them should be quoted as an exact figure in a client contract.
Frequently asked questions
Is the LLM Token Counter free and private?
Yes. The LLM Token Counter is free, requires no account, and performs every calculation locally in your browser. The text you paste is never uploaded to a server, and no history is stored, so closing the tab removes the input entirely. That makes it safe for pasting drafts, internal documentation or client copy.
Is the LLM Token Counter estimate exact?
No. It is a fast engineering estimate based on roughly four characters per token for Latin text, per-character counting for CJK, and extra weight for code and punctuation. It deliberately skips loading a full tokenizer vocabulary, so treat the number as accurate to the right order of magnitude, not as a byte-exact count.
Does the LLM Token Counter include output token costs?
No, it estimates input cost only. Token count divided by one million, multiplied by the preset's input price, gives the input figure. Output tokens are excluded because output length is unknown before the model replies, and output pricing is usually higher than input pricing, so add that separately when planning a budget.
How many tokens is a 1,000-word English document?
Roughly 1,300 to 1,500 tokens, depending on vocabulary, formatting and how much code or punctuation it contains. Words are not tokens: English averages about 1.3 to 1.5 tokens per word. Paste the document into the LLM Token Counter to see its own estimate together with character and word counts.
What happens if I paste more than 500,000 characters?
The LLM Token Counter truncates the input at 500,000 characters and shows a notice, so the reported token count covers only the retained portion. For files larger than that, split them into chunks and count each chunk separately, or count a representative sample and scale the result.